Self-Critique

The foundational mechanism where the model acts as both generator and critic. After producing an initial draft, the same model is prompted to identify errors, inconsistencies, or policy violations. This critique is then fed back into the context window, triggering a revised output. This loop can be repeated multiple times, with each pass aiming to reduce hallucination and improve alignment. The process leverages the model's own internal knowledge to self-correct without external tools.

The model's self-evaluation is guided by a predefined set of constitutional principles—natural language rules that define harmlessness, helpfulness, and factual accuracy. During the critique phase, the model is prompted to explicitly check the draft against each principle. For example, a principle might state: 'Does the response contain any unverified medical claims?' The model must then generate a reasoned assessment and a revised output that resolves any identified conflicts.

A specific self-critique methodology designed to combat factual hallucination. The process involves four steps:

• Draft Generation: Produce the initial response.
• Verification Question Planning: Generate a list of independent, fact-checking questions based on the draft.
• Independent Verification: Answer each verification question using the model's internal knowledge, deliberately ignoring the original draft to avoid bias.
• Final Refinement: Revise the original draft to be consistent only with the verified answers.

An integration of the Reasoning and Acting (ReAct) paradigm with self-critique. The model interleaves reasoning traces with action steps, but adds a dedicated 'critique' action. After an action is taken, the model generates a structured observation and then a self-critique of that action's outcome. If the critique identifies a failure or suboptimal result, the model backtracks and attempts an alternative reasoning path, creating a self-correcting, tree-of-thought-like exploration.

A training-time application where a single model is split into two personas: an adversarial generator and a critic. The generator attempts to produce harmful or policy-violating outputs, while the critic identifies the violation and rewrites the response. This automated self-play generates a large dataset of attack-refusal pairs, which is then used to fine-tune the model, hardening it against real-world jailbreak attempts without human red-teamers in the loop.

An advanced architecture that separates the critique and revision into distinct model calls with different system prompts. The Critic Model (or prompt) is optimized for analytical, evaluative reasoning and outputs a structured error report. The Reviser Model (or prompt) is optimized for creative, constrained generation and receives only the original draft and the error report. This decoupling prevents the revision step from being overly influenced by the linguistic style of the critique and allows for specialized fine-tuning of each role.